Well-Formed Dependency to String translation with BTG Grammar

Well-Formed Dependency to String translation with BTG Grammar Xiaoqing Li†, Kun Wang†, Dakun Zhang‡, Jie Hao‡ † National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences ‡ Toshiba (China) R&D Center {xqli,kunwang}@nlpr.ia.ac.cn {zhangdakun,haojie}@toshiba.com.cn Abstract This paper proposes a well-formed depen- dency to string translation model with BTG grammar. By enabling the usage of well- formed sub-structures and allowing flexible reordering of them, our approach is effective to relieve the problems of parsing error and flatness in dependency structure. To utilize the well-formed dependency rules during decod- ing, we adapt the tree traversal decoding algo- rithm into a bottom-up CKY algorithm. And a lexicalized reordering model is used to en- courage the proper combination of two neigh- bouring blocks. Experiment results demon- strate that our approach can effectively im- prove the performance by more than 2 BLEU score over the baseline. 1 Introduction Due to the merits of holding shallow semantic infor- mation and cross-lingual consistency (Fox, 2002), dependency grammar has attracted much attention in the field of machine translation (Lin, 2004; Quirk et al., 2005; Ding and Palmer, 2005; Shen et al., 2008; Xie et al., 2011). The dependency-to-string model (Xie et al., 2011) falls into the paradigm of ”translation after under- standing”, which tries to understand the structure and meaning of source text. However, there are two typical problems for this approach. One is that the model is prone to be affected by parsing errors. Dependency-to-string model adopts a unique source side tree structure as fixed input and constructs the output by converting each sub-structure into target side. If there is some errors in the source depen- dency tree, for example, a prepositional subtree is attached to a wrong head, the model can hardly re- cover the error. Figure 1(a) shows another parsing error, in the correct parsing result, ”与北韩...国家” should forms a subtree with ”国家” as the head, and then this subtree is dominant by ”之一”. The other problem is that dependency structure is too flat for the translation task. Since dependency- to-string model requires a head and all its depen- dents to be translated as a whole, the flatness of the structure will make rules difficult to be matched dur- ing decoding. Furthermore, it will also lower the ro- bustness of translation rules, since many giant and low-frequency rules will be extracted. Figure 1(b) shows an example of the flat structure. The head word ”提供” has five dependents in the structure. If no rule can be matched, only the glue rule can be applied. As a result, the prepositional subtree ”为 了...义务” cannot be correctly reordered to the end in the target side. Existing solutions for the above problems include (Meng et al., 2013; Xie et al., 2014). The form- mer incorporates phrasal nodes of constituency trees in the source side of translation rules, and the lat- ter modifies translation rules during decoding to al- low the usage of phrases which are compatible with well-formed structures. Since these two approaches still adopt head-dependent structure …

Content truncated to 500 words